FASID Discussion Paper Series on International Development Strategies No. 2003-006

WORKING-AGE ADULT MORTALITY AND PRIMARY SCHOOL ATTENDANCE IN RURAL KENYA

Takashi Yamano Foundation for Advanced Studies on International Development, Japan

T. S. Jayne

Michigan State University, USA

November 2003

Foundation for Advanced Studies on International Development

WORKING-AGE ADULT MORTALITY AND PRIMARY SCHOOL ATTENDANCE IN RURAL KENYA

Takashi Yamano and T. S. Jayne

Abstract

The rapid increase in adult mortality due to the AIDS epidemic in sub-Saharan Africa raises great concern about its impact on child welfare. This article estimates the impact of AIDS-related adult mortality on primary school attendance in rural Kenya using a panel of 1,266 households surveyed in 1997, 2000, and 2002. We find a strong correlation between working-age adult mortality and lagged HIV-prevalence rates at nearby sentinel survey sites. School attendance, especially for children in relatively poor households, is negatively correlated with lagged provincial HIV-prevalence rates. Children, especially girls in relatively poor households, are less likely to be in school directly prior to the death of an adult member than children in unafflicted households. By contrast, boys in relatively poor households are less likely to be in school after an adult death. The evidence indicates that rising adult mortality in rural Kenya is adversely affecting primary school attendance especially among the poor. However, these results measure only short-term impacts. Over the longer run, whether school attendance in afflicted household rebounds or deteriorates further is unknown. JEL classification: O12, O15, J10, Q12 Keywords: HIV/AIDS, Education, Kenya

1

1. INTRODUCTION

Education contributes to economic growth (Barro, 1991; Mankiew, Romer, and Weil,

1992). Rising education of the labor force has been identified as a primary ingredient in processes

of structural transformation and economic development (Johnston and Kilby, 1975;

Gabre-Madhin and Johnston, 2002). Education is a primary investment that the poor can make to

escape from poverty over the long run. Recognizing this, many governments and international

organizations put a high priority on education in their development agendas, such as the

Millennium Development Goals (United Nations, 2000).

The onset of the AIDS epidemic in the past two decades raises serious questions about long

term human capital development, including education, and the future quality of the workforce in

areas where the disease is particularly acute, such as in many countries of Africa. There is growing

concern over the effect of high AIDS-related adult mortality and illness on child welfare and the

disease’s effects on their potential over the long run to support themselves as adults and to

contribute to their countries’ development (UNICEF, 1999; Bell, Devarajan, and Gersbach, 2003).

Available empirical evidence indicates that school enrollment rates are lower among AIDS

orphans compared with non-orphans (World Bank, 1999; Ainsworth, Beegle, and Koda, 2002;

Case, Paxson, and Ableidinger, 2002). A long list of anecdotal case studies also supports this

evidence (Gachuhi, 1999; Nyambedha, et al., 2001; Guest, 2001; USAID, 2002). However, using

39 nationally representative data sets collected in the 1990s from 28 countries, mostly in

Sub-Saharan Africa, Ainsworth and Filmer (2002) show that the difference in enrollment rates

between orphaned and non-orphaned children varies greatly across countries and wealth levels

within a country. There are likely to be important conditioning factors affecting the relationship

between parent mortality and child schooling. 2

A major difficulty in measuring the impact of adult mortality, especially mortality

attributable to AIDS, is that it is caused by behavioral choices rather than by random events.

Individuals and households incurring adult mortality are more likely to display certain

characteristics. For example, especially in the early years of the epidemic in sub-Saharan Africa,

evidence suggests that men and women with higher education and income were more likely to

contract HIV than others because they were more able to attract potential sexual partners

(Ainsworth and Semali, 1998; Gregson, Waddell, and Chandiwana, 2001).1 If prime-age

mortality remains correlated with individual and household characteristics such as social status,

wealth, and mobility – which are also important determinants of school enrollment – failure to

control for these characteristics may generate biased estimates of the impact of adult mortality on

school attendance. However, it is possible to overcome this problem by estimating the effect of

adult mortality on school attendance of the same children over time, using child fixed-effects

models. To our knowledge, no study has yet applied child fixed-effects models to overcome this

problem.

The purpose of this paper, therefore, is to identify the effect of adult mortality on school

attendance among primary-school age children (7 to 14 years) in rural Kenya. To address this

question, we first identify potential pathways by which adult mortality may affect child school

attendance and the timing of each pathway. Then, we estimate the impacts of adult mortality on

child school attendance in rural Kenya after stratifying sampled households and their children by

1 As information about HIV transmission spreads, however, it is believed that educated people are

more likely to change their behavior in ways that reduce their vulnerability to the disease

compared to less educated people.

3

wealth and gender. Kenya is one of the most heavily HIV-infected countries in the world: 13.5

percent of adults aged 15 to 49 are estimated to be living with HIV in June 2000 (NASCOP, 2001).

Two sources of data are used: a three-year nationwide panel data set of 1,422 rural

households, collected in 1997, 2000, and 2002, and secondary data on HIV-prevalence rates from

13 sentinel surveillance sites, collected annually between 1990 and 1999 by the National AIDS

and STDs Control Programme (NASCOP). The data are used to estimate the effect of

working-age adult mortality since the first survey on school attendance of children aged 7 to 14 in

the second and third surveys.

The study highlights three major findings: First, children’s school attendance is adversely

affected by the death of working-age adults among the bottom half of sample households ranked

by initial asset levels in 1997, but no significant effects are detected among households in the top

half of the asset distribution. Second, working-age adult mortality negatively affects school

attendance even before the death in poor households. The negative impact is larger among girls

than boys. This suggests that children, especially girls, are sharing the burden of caring for sick

working-age adults and/or that school fees tend to be among the first expenditures curtailed in

relatively poor households after one of their prime-age members becomes chronically ill. By

contrast, school attendance among boys in relatively poor households drops more sharply after the

adult member dies. Third, school attendance among the poor is negatively correlated with lagged

provincial HIV-prevalence rates, even after controlling for child fixed-effects. This result

suggests that AIDS is indirectly affecting child school attendance in ways other than through the

death of household members, or that unobserved time varying factors correlated with changes in

HIV-prevalence rates are impeding school attendance.

4

The next section of the paper describes the sample and panel data used in the analysis.

Section 3 presents the conceptual framework and some descriptive analyses. Estimation strategies

and variables are discussed in Section 4. Estimation results are in Section 5, followed by

conclusions in Section 6.

2. DATA

2.1. Working-age Adult Mortality

The study uses a three-year panel of rural household surveys in 1997, 2000, and 2002. In

April 1997, a total of 1,578 households were randomly selected from 24 districts within the eight

agriculturally-oriented provinces of the country. The sampling frame for the surveys was prepared

in consultation with the Central Bureau of Statistics and implemented by the Tegemeo Institute of

Egerton University. We exclude two pastoral districts (40 households) that differ substantially

from other zones and had high rates of attrition. Thirty-eight households having farm sizes greater

than 20 acres were also excluded, to maintain the study’s focus on small-scale households. Of the

1,500 remaining households, 1,422 households were able to be re-interviewed in 2000, and of

those, 1,266 households were re-interviewed in 2002 (Table 1).2

When enumerators re-visited the 1997 sampled households in 2000 and 2002, they asked

the whereabouts of each individual in the demographic roster of the 1997 survey. If the household

contained new members not listed in the 1997 survey, the enumerators added the new members’

names in the demographic roster, obtained socio-demographic information on these individuals,

2 Please see Yamano and Jayne (2003) for detailed discussions on the sample and analyses on the

impacts of adult mortality on household composition, farm production, and assets.

5

and included them in the roster for subsequent surveys. Similar information was recorded for

members in 1997 who were no longer residents of the household in 2000 or 2002. Through this

process, information was obtained on individuals who had passed away since the preceding

survey.

Our analysis focuses on adults between the years of 15 to 54 for men and 15 to 49 for

women because these correspond to the age ranges reported by the World Health Organization for

HIV prevalence, and are the age ranges hardest hit by the disease (NASCOP, 2001). Between the

1997 and 2000 surveys, 84 adults in these age ranges (which we hereafter refer to as

“working-age” adults for shorthand) passed away among the 1,422 sampled households. In the

2002 survey, we found 42 additional cases of mortality in these age ranges among the 1,266

households since the 2000 survey. Thus, the total number of deceased working-age adult members

is 126. Of those, about 36 percent were found in Nyanza province, where only 12 percent of the

sampled households reside.

In Nyanza province, the survey covered three districts: Kisumu, Siaya, and Kisii districts.

In Kisumu and Siaya districts, 23.7 percent of the sampled households experienced at least one

working-age adult death since the 1997 survey. In Kisii district, which is farther away from the

main highways, more rural, and culturally different from the Luo dominated areas of Kisumu and

Siaya, only 4.7 percent of sampled households had experienced at least one working-age death

since the 1997 survey (Table 1, column C). These differences in death rates across districts are

broadly consistent with the differences in HIV-prevalence among pregnant women testing

HIV-positive at urban sentinel surveillance sites in these district (Table 1, column F).

In Kisumu city, the third largest city in Kenya and the capital of Nyanza province, about 22 percent

of pregnant women tested HIV positive at a sentinel-surveillance site from 1990-94. The five-year 6

period of 1990-94 is matched with the survey data over the 1997-2002 period to reflect the 6-10

year average survival time after HIV sero-conversion as noted by epidemiological studies in the

region (Whitworth et al., 2003; Todd, 2003).

Kisumu and Siaya districts are well connected to Kisumu city with busy highways that also

connect Kisumu with Nairobi to the east and with Kampala, Uganda, to the west. Along highways

in Kenya, especially the Trans-Africa Highway that connects Mombasa with Kampala via Nairobi,

high HIV-seroprevalence rates have been found among high-risk people, e.g., truck drivers and

commercial sex workers (Carswell, Lloyd, and Howells, 1985; Mbugua et al., 1995). HIV appears

to spread into rural communities from truck stops via interactions between high-risk people and

local people, including adolescents who interact with high-risk people (Nzyuko, et al., 1997).

Siaya district is also connected with Busia, a border city with Uganda, where about 20 percent of

pregnant women were found to be HIV positive at a sentinel surveillance site during 1990-94

(NASCOP, 2001).

Additional factors are believed to have contributed to a high HIV prevalence rate in

Kisumu and Siaya districts. Such factors include the relative importance of polygamous

households (which increases HIV transmission between a husband and wives), Luo peoples’

traditional practices, such as widow inheritance (which increases HIV transmission from a widow

to a new husband) and customs associated with fish trading (where sexual favors are common

among fishermen and women traders), and male non-circumcision (which increases men’s risk of

HIV infection per sexual contact).

By contrast, the district town of Kisii, which is further from the Trans-Africa Highway and

home to the Kisii ethnic group, only 3.6 percent of pregnant women tested HIV positive at a

sentinel surveillance site from 1990 to 1994 (NASCOP, 2001). Thus, the difference in the 7

numbers of working-age adult deaths between Kisumu/Siaya districts and Kisii district is

consistent with the difference between the HIV prevalence rates between districts. To examine

this point further, we estimate the association between lagged HIV-prevalence rate at the nearest

urban-sentinel-surveillance site and the probability of experiencing working-age adult death in

Section 5.

2.2. Primary School-Age Children Included in the Analyses

In the 2002 survey, respondents were asked about each household member’s schooling,

including the number of years spent in school prior to the survey and whether each child was

currently attending school. The results on school attendance in columns D and E of Table 1 are

taken from 2,565 and 2,107 children aged 7 to 14, who had yet to complete their primary school

education by the time of the 2000 and 2002 surveys, respectively. This sample includes children

who had left the sampled households since the 1997 survey. Excluding them from the final sample

may cause an under-estimation of the negative impact of adult mortality on school attendance

because it is more likely that children who left (or were sent away from) their households after

experiencing adult mortality were less likely to be attending school than those who remained in the

sampled households. In the 2000 and 2002 survey, enumerators asked about the schooling of

children listed on the 1997 survey even when the children had left the sampled households.

Among the 2,565 and 2,107 children in the 2000 and 2002 sample, 8.3 and 11.8 percent of children

left their households (either permanently or temporarily) since the 1997 survey, respectively.

Second, despite our efforts to keep track of all of the sampled households and individuals,

we lost some children in the final samples due to attrition. For instance, between the 2000 and

2002 surveys, we lost track of 345 children aged 7 to 14 years old. Out of the 345 children, 259 8

children are in the 156 households that could not be re-interviewed in 2002. If these households

suffered a higher incidence of working-age mortality between 1997 and 2002, we would have

sample selection bias when estimating the impact of working-age adult mortality on child

schooling. To examine this attrition bias, we include an attrition dummy variable in the estimation

models.

Finally, to maintain our focus on primary school enrollment, we exclude from the final

sample 28 and 96 children who were between 7 and 14 years old in 2000 and 2002, respectively,

because respondents indicated that they had already completed primary school prior to the

surveys.

3. Impact of Prime-age Adult Mortality on Schooling

3.1. Conceptual Framework

How does working-age mortality reduce child schooling? There are three main economic

factors influencing child school enrollment: (a) the financial costs of schooling, such as school

fees and books, relative to households’ resources; (b) the opportunity costs of children’ time; and

(c) the expected returns from school. The potential effects of working-age mortality on child

schooling depends on how working-age mortality affects these factors (World Bank, 1999).

First, medical expenditure associated with prolonged illness and eventual funeral costs

reduce the financial resources of the household (Barnett and Blaikie, 1992; Lundberg, Over, and

Mujinga, 2000). Depending on the household’s initial income and wealth and impact of death on

these variables, the household may withdraw children from school. Because afflicted households’

financial situation may not recover quickly after the death of the sick member, children’s school

attendance may be affected long after the death occurs. 9

Second, AIDS-related adult mortality is likely to negatively affect child schooling by

increasing the opportunity costs of children’s time. As one family member becomes chronically ill,

another household member (usually a female member and often an older girl) must devote more

time to care giving.3 As a result, a care-giving female member may need to reduce time devoted

to her usual activities. If a girl is expected to help care for the sick, she may stay home instead of

going to school. The increased demand for care-giving labor, however, will disappear if the sick

adult passes away.

The household also experiences an increased demand for labor previously provided by the

sick member. The way in which households adjust to internal labor supply shocks varies

according to the resources of the households. For example, households with sufficient income

may hire additional workers to meet residual labor needs. Some afflicted households are able to

attract additional members to at partially offset the loss of another member (Beegle, 2003;

Ainsworth, Ghosh, and Semali, 1995). Households that are poor and/or face high opportunity

costs on their scarce resources may be most vulnerable to calling upon their children to discontinue

school to provide labor to the household after an adult passes away. An increased demand for

labor among non-sick members will not disappear even after the death of the sick member.

Third, because of its effect on life expectancy, HIV/AIDS in high-prevalence countries

may alter individuals’ time preference for money (McPherson, 2001). Life expectancy in Kenya

3 For instance, Ainsworth and Dayton (2003) found a decline in Body Mass Index among elderly

in non-poor households after an adult death. They suggested that the decline could be explained

by a diversion of household resources to the medical care and/or increased demands on the time

of elderly to care for sick AIDS patients in non-poor households.

10

in 1999, for instance, is 48 years, compared to 58 years in 1993 (World Bank, 1995; 2002). Part

of the decline in life expectancy is due to increased infant mortality caused by mother-to-child

HIV infection, yet there is still a considerable decline in the expected years of work among young

adults. The decline in the expected years of work is expected to change parents’ expectations

about the lifetime return from their children’s education both to themselves (e.g., intergenerational

transfers from children to parents in their old age) and to their children.

We lack suitable instruments to empirically distinguish the impacts of one factor from the

others. However, this conceptual framework provides some insight into the potential pathways by

which adult mortality may affect child schooling. We consider two hypotheses in particular. The

first hypothesis is that children in poor households are more vulnerable to working-age adult

mortality because of the first and second factors. To test this hypothesis we stratify the sample into

two groups based on households’ initial asset levels and compare model results for both groups.

According to this hypothesis, we expect to find larger impacts of working-age adult mortality

among the relatively poor half of the sample than the less-poor half.

The second hypothesis is that adult mortality affects boys and girls differently. In Kenya,

as in most parts of eastern and Southern Africa, girls tend to assume more care giving

responsibilities than boys (Opiyo, 2001). Thus, if the second pathway is important, we expect to

find a larger negative impact on girls’ schooling than that of boys prior to adult mortality. Thus,

we also stratify the sample by gender.

AIDS may affect child schooling not only through direct impacts on afflicted households

but also through community effects. For instance, areas with high HIV prevalence rates suffer

reduced profitability of businesses such as commercial farms, outgrower schemes, and non-farm

businesses relying on wage labor (e.g., Fox et al., 2003; Rugalema, 1999). Reduced economic 11

profitability may translate into wage job contraction, fewer economic opportunities, and lower

household incomes in such areas, which may indirectly affect children's school attendance. For

these reasons, studies measuring the effects of AIDS on child schooling only through the pathway

of how afflicted households themselves respond may underestimate the actual impact of the

disease.

3.2. Descriptive Analysis

Based on data collected from the 2000 and 2002 surveys, we categorize children in the

sample according to whether they incurred the death of a working age adult in their household in

either of the two survey intervals: September 1997- May 2000 and May 2000 – May 2002. In the

following discussion, we call the interval between the 1997 and 2000 surveys as the first period

and the interval between the 2000 and 2002 surveys as the second period.

In Table 2, we group children by whether they suffered the death of a working-age adult in

their household in either the first or second period. Among these different groups of children,

Table 2 shows differences in school attendance rates, years of schooling, age, and numbers of

sample children in each category.

Overall mean attendance rates4 are 89.7 and 88.3 percent in 2000 and 2002, respectively.

But among children who experienced working-age adult mortality in the first period (Table 2,

column C and D), the mean attendance rate declines from 91.1 percent in 2000 to 81.6 percent in

2002. Thus, we find a larger decline in enrollment rates among children who experienced the

4 Throughout this paper, we use the “attendance” to signify whether the child was actually in

school, as opposed to “enrollment,” which does not necessarily imply regular presence in school.

12

mortality than children who did not. This suggests that children are dropping out of school, after

experiencing adult mortality in their households. When we stratify the sample into two groups

ranked by initial value of productive assets – hereafter, poor and non-poor -- we find a larger

decline in enrollment among the poor (from 86.4 to 72.1) than non-poor (from 94.5 to 86.6).

On the other hand, children who experienced working-age adult mortality in the second

period (between 2000 and 2002) were less likely to be in school in 2000 (i.e., before the death

occurred) than in 2002 (Table 2, column E and F). The average attendance rate jumps from 70.0

percent in 2000 to 89.6 percent in 2002, after experiencing the adult mortality. This increase in

attendance is much more drastic among the poor than non-poor. Low attendance rates in 2000

may reflect a combination of the added burden on children to care for the ill member in the first

period, prior to the adult mortality, and/or to provide labor activities formerly performed by the ill

member.

Comparing the years of schooling completed by children, we note that children in poor

households who incurred a working-age adult mortality during the second period have

significantly less schooling, 1.90 years, in 2002 (Column F) than children who did not experience

any working-age adult mortality, 2.66 years, (Column B), although their average ages are very

similar. This suggests that children who experienced working-age mortality in the second period

had delayed enrollment or dropped out temporarily from school, possibly to take care of ill

members. This could explain why we find a relatively low attendance rate in 2000 but a high

attendance rate in 2002 among children who experienced working-age mortality in between 2000

and 2002 (the second period).

The findings in Table 2 suggest a process described in Figure 1. When an adult starts to

become ill from AIDS, the probability drops that a school-age child will attend school. Because 13

medical expenses tend to rise at the same time that earnings from the sick member declines, we

expect that adjustments in school enrollment will be greater for relatively poor households. After

the death of the sick member, the child’s likelihood of being in school rises because care giving

needs have subsided. However, depending on whether the household has been able to attract

additional labor to take over tasks formerly handled by the deceased, the opportunity cost of

children’s labor may still be perceived to be too high to warrant an immediate return to school.

Relatively poor afflicted households may also be less able to continue incurring school fees. By

contrast, children in better-off households are less subject to a dramatic “down-up-down”

movement, because of their households’ greater ability to draw upon savings and assets to

maintain desired expenditures in the face of economic shocks.

The picture emerging from Table 2 and Figure 1 are only bivariate associations, which may

be spuriously driven by regional differences or household characteristics. Thus, to measure the

differences in enrollment rates due to adult mortality, holding other factors constant, we need to

employ multivariate techniques. In addition, because HIV infection is influenced by behavioral

choices of household members, the possible correlation between working-age adult mortality and

unobserved factors in schooling also needs to be controlled for.

4. Estimation Strategies and Variables

Our estimation procedure followed four steps described below. We first develop a working-age

mortality model to identify factors associated with children experiencing the death of a

working-age member in their household over the entire survey period. Next, we estimate primary

school attendance models to examine the effects of mortality on schooling both before and after

the timing of the death. Lastly, to examine the robustness of the findings after controlling for 14

time-invariant unobserved effects, we estimate household- and child- fixed effects conditional

logit models of child school attendance.

4.1. Estimation Strategies

The Working-Age Adult Mortality Model

Before estimating the impact of working-age adult mortality on schooling, we first

estimate the determinants of children’s experience of working-age adult mortality in the first

period (1997-2000). We focus on the first period because the sample in the first period has a less

severe attrition problem than the second period. To examine a relationship between the

lagged-HIV prevalence rate and sample attrition, we include a dummy variable for attrition, Ait.

The attrition variable, Ait, equals one for children whose information is not available in 2002. Thus,

the working-age adult mortality model is

Di1997-2000 = φHIV HIVj 1990-94 + φA Ait + φX Xit + e it (1)

where Di1997-2000 is a dummy variable that takes one if child i experienced working-age adult

morality in 1997-2000; HIVj 1990-94 is the average lagged-HIV-prevalence rate in at the nearest

surveillance site in 1990-94; and Xit is a vector of child, household, and regional characteristics.

We estimate this model with Probit. Because of a long asymptomatic period, we include the

lagged-HIV prevalence information from the period of 1990-1994. By estimating φHIV and

establishing the relationship between working-age adult mortality and the lagged HIV-prevalence

rate, we can examine the extent to which working-age mortality found in the panel data are

influenced by HIV/AIDS.

The Primary-School Enrollment Model 15

The economic theory and estimation models on schooling have been discussed in

numerous papers (Strauss and Thomas, 1995; Glewwe, 2002). Following previous studies, we

write a simple equation of a school enrollment model as:

Sit = βHIV HIVjt-K + βD Dit + βX Xit + u it (2)

where Sit is a dummy which takes one if child i is enrolled in school at time t; other variables are

as defined as before. The lagged-HIV prevalence rate at the nearest surveillance site is expected

to pick up broader community effects of the AIDS epidemic on child school attendance (separate

from the direct effect via afflicted households). However, the lagged HIV-prevalence rate could

be correlated with many regional characteristics. For instance, we know from previous studies

that HIV prevalence rates tend to be high in areas with major trunk roads where there is a steady

influx of outsiders. Thus, we should be cautious when we interpret the results on this variable.

We match the lagged mean HIV-prevalence information from 1990-1994 (1992-1996) with the

observations of the 2000 (2002) survey because of the 6-10 year average survival time after HIV

sero-conversion (Whitworth et al., 2003; Todd, 2003).

Before proceeding, we clarify the notation for survey periods. In the first period (1997 to

2000), which we call t=1, the schooling dummy (Si1) indicates whether child i is in school at the

time of the 2000 survey. Analogously, Si2 indicates whether child i is in school at the time of the

2002 survey. The working-age adult mortality dummy, Di1 (Di2) indicates whether child i resides

in a household experiencing at least one working-age death in the period between the 1997 and

2000 (2000 and 2002) surveys.

We also want to ascertain whether children’s school attendance is affected prior to and/or

after the occurrence of a death. This is because the impact of the adult mortality may manifest long

before the member actually dies, as discussed in Section 2. We therefore include two categorical 16

variables in the models to account for time of death relative to the observations on child schooling.

The post-mortality variable (Dit-1), indicates that child i’s household experiences a working-age

death in the period prior to the observation on schooling. For instance, if a child experiences a

working-age death in the household in the first period, the post mortality variable (Dit-1) takes a

value of one in the second period and zero in the first period. The pre-mortality dummy variable,

on the other hand, indicates that child i experiences a working-age adult death in the next period.

For instance, if a child experiences a working-age death in the household in the second period,

then the pre-mortality variable (Dit+1) takes a value of one in the first period and zero in the second

period.

Finally, we also include the attrition dummy variable in Equation 2 to test whether the

sample attrition problem may cause an attrition bias in model estimation. If the attrition variable

is not significant, then the result indicates that the children whose information is missing in 2002

are not statistically different from other children, after controlling for observed information. This

test is proposed by Fitzgerald, Gottschalk, and Moffitt (1998) and used by others, e.g., Alderman

et al. (2001).

Thus, the estimated schooling equation is

Sit = βHIV HIVjt-K + βD Dit + βpost Dit-1 + βpre Dit+1 + βA Ait + βX Xit + u it (3)

We use pooled data from the 2000 and 2002 surveys in this probit estimation. As discussed in

Section 2, the sample is stratified into two groups ranked by initial wealth in 1997, and by gender,

to obtain better understandings on the pathways that the adult mortality affects the child schooling.

Lastly, we re-estimate Equation 3 using fixed effects. Household fixed effects models

purge the potential correlation between working-age adult mortality variables and unobserved

household-fixed effects from the estimates while still being able to estimate the effects of 17

child-specific characteristics, such as relationship to the household head. Further, we can purge

child-fixed effects from the estimates by using the child-fixed effects model. We estimate these

fixed effects models with conditional logit models.

When estimating Equation 3 with the household- or child-fixed effects model, we cannot

include time-invariant variables. In addition, we need to exclude one of the three working-age

adult mortality variables (pre-, during-, and post-) because we have only two observations for each

child. Therefore, we exclude Dit from Equation 3 when we estimate the household- or child-fixed

effects model. The pre- and post-mortality variables in the child-fixed effects model measure the

difference in probabilities of the same child being in school depending on whether he or she

experiences the working-age adult mortality before or after the observation on school.

4.2. Variables

As discussed in Section 2, we have excluded children who had already completed primary

school. Thus the dependent variable, Sit, is an indicator of whether a child is attending a primary

school who is eligible for enrollment. The dummy variable on working-age mortality, D t, equals

one if there is at least one working-age adult death in a household during a period t. The

HIV-prevalence variable, HIVj t-K, is the average ratio of pregnant women who visited the nearest

urban sentinel surveillance site and tested HIV positive during the five-year period from 1990-94

for the 2000 samples and from 1992-96 for the 2002 samples. Data from 11 such sites, maintained

by the National AIDS and STDs Control Programme (NASCOP), were matched at the

district-level to all households in the sample (Appendix Table A2). While unobserved time

invariant effects are controlled for using child fixed-effects models, it is still possible that the

lagged HIV-prevalence term may be picking up the effects of time variant unobservables. With

18

this caveat, we include lagged HIV-prevalence rates in the models to account for broader

community level effects of AIDS on child schooling.

We also include an attrition variable equalling one for 345 children whose information is

not available in 2002 either because their households were not interviewed in the 2002 survey (75

percent of the 345 children) or they could not be identified by the respondents in their households

(25 percent). The descriptive statistics of these and other variables are presented in Appendix

Table A1.

Child characteristics include the age of the child, its squared term, and a gender dummy

variable for girls. We also include two dummy variables indicating whether the child’s

relationship to the household head is as a grandchild, nephew, or niece, or as a distant relative or

a non-relative who lives in the household.

Because the effects of working-age mortality are potentially devastating, household

characteristics that are usually considered exogenous or fixed in most household models could be

rapidly and severely affected by adult mortality. Because of this, we feel it is inappropriate to

include in the models current values of variables such as asset levels, family size, or landholding

size. For this reason, we use household characteristics based on values from the initial 1997

survey. These household characteristics are the years of schooling for the most educated person in

the household in 1997; a binary variable for polygamous households; a binary variable for

female-headed households in 1997, the acres of land owned in 1997 in logs; and the value of asset

holdings in 1997 in logs.

We also include two community variables designed to control (albeit imperfectly) for access to

markets and services, and interactions outside the village community: distance of the village to the

nearest bus/taxi stop, and distance to the nearest piped water outlet. Finally, five provincial 19

dummies are included, with Eastern Province as the reference province.

5. Results

5.1. Determinants of Experiencing Prime-age Adult Mortality

We first discuss the results from Equation 1, the determinants of the probability of

experiencing at least one working-age adult death in the first and second periods. The binary

dependent variable, D1997-2000, equals one if a child resides in a household experiencing at least one

working-age adult death over this three-year period. Table 3, column A presents the results

without including the lagged-HIV prevalence variable.5 As expected from the results in Table 1,

children in Nyanza province are 8.5 percent more likely to experience working-age adult mortality

in their households than children in Eastern province (the reference group). Once the lagged

HIV-prevalence variable is included in column B, however, the estimated coefficient of Nyanza

province becomes smaller. This provides evidence that working-age adult mortality in Nyanza

province is strongly associated with HIV (and thus AIDS). The lagged-HIV prevalence variable

is highly correlated with a child experiencing adult death in his/her household at the one percent

level of significance.6

5 We report the marginal probabilities in Table 3 and the following tables. For dummy variables,

the “marginal” probabilities are calculated by changing the value of dummy variables from zero to

one, holding other variables constant at their sample means.

6 When we tried the current HIV prevalence rate from 1995-99 instead of the lagged HIV

prevalence rate from 1990-94, the size of the estimated coefficient declined to 0.321 (t-stat =5.19)

from 0.487 (t-stat=5.22). Thus, the lagged HIV prevalence rate is better correlated with the

20

On the attrition variable, we find a positive correlation between the attrition variable and

the probability of experiencing the adult mortality. Thus, this indicates that the attrition did not

occur randomly and suggest a possibility of an attrition bias in the next model where we use the

pooled data from the 2000 and 2002 surveys. As far as the results in Table 3 are concerned, we do

not need to worry about the attrition bias because we only use the cross-section data from the 2000

survey, before the attrition occurs.

Turning to the other variables in the model, we find that children who are away from bus

stops are less likely to experience adult mortality in their households. This is consistent with

previous studies that HIV-prevalence rates are relatively high along the major highways, as

discussed in section 2.2.

We also find that children in polygamous households have a lower probability of

experiencing working-age adult mortality than children in monogamous households. This could

be because we do not count polygamous husbands as household members unless the husbands

actually live in the sampled household for more than six months. On the other hand, an additional

male member in the household increases the probability of children experiencing working-age

adult mortality by 1.4 percent. By contrast, the number of female members does not significantly

affect the probability. Lastly, a household’s landholding size (in acres) positively influences the

probability of experiencing working-age adult mortality. This suggests that children who

experienced working-age adult mortality in the first period tend to be from households with

greater assets and socioeconomic status.

probability of experiencing the adult mortality than the current HIV prevalence rate as the

long-asymptomatic period suggests.

21

Overall, these results suggest a grave situation for children in rural Kenya in the near

future because HIV-prevalence rates among pregnant women at the sentinel surveillance sites

have increased between 1990 and 2000 in nine of the eleven urban-sentinel-surveillance sites

nationwide (NASCOP, 2001). A simple simulation based on the results in Table 3 indicates that

an increase in the HIV-prevalence rate from 5.0 percent to 13.5 percent (from the nationwide level

in 1990 to that in 2000) would increase the probability that a child experiences the death of a

working-age adult over next five years from 2.39 percent to 6.07 percent.7 Thus, unless there is

significant progress in reducing HIV prevalence rates, there will be a significant increase in the

number of rural children in Kenya faced with the death of a working age adult in their households.

5.2. Impact of Working-age Adult Mortality

In Table 4 and 5, we present the Probit results based on Equation 3. First, we estimate the

model for all children. Second, we stratified the sample into two groups based on the initial wealth

in 1997. Finally, we stratified the sample by gender (Table 5).

The HIV prevalence rate

The results in Table 4 indicate that the HIV prevalence rate at the nearest sentinel site

negatively influences children’s schooling among the poor (children in households whose total

asset value was below the median in 1997). A simple simulation based on the results among the

7 These probabilities are calculated after changing the value of the HIV prevalence variable from

one value to another, using mean sample values of all other variables except the HIV prevalence.

22

poor in Table 4 (column B) indicates that the probability of attendance decreases by 2.5 percentage

points from 92.1 percent to 89.6 percent when the HIV prevalence rate increases from 5.0 percent

to 13.5 percent (from the 1990 level to the 2000 level). We find no significant impact of the HIV

prevalence rate on child schooling among the relatively less poor.

The HIV prevalence rate may be capturing the indirect effects of the HIV/AIDS epidemic,

other than the direct household-level effects of adult mortality on schooling, because we already

include information on working-age adult mortality in the model. However, it is also possible that

the HIV prevalence rate and child school attendance are correlated through other regional factors.

The working-age adult mortality

Working-age adult mortality does not significantly influence children’s schooling at the

time of death. But it negatively affects children’s schooling before and after they experience

working-age adult mortality among the poor (column B in Table 4). When we stratify the sample

by gender, we find a larger negative impact of the pre-mortality variable among girls than boys in

poor households (column C and D in Table 5). This indicates that girls are likely to be taking care

of the sick members in the households prior to their deaths. The result indicates that girls in the

poor households are 15.1 percent less likely to be in school in a period prior to the adult mortality,

while boys are 12.4 percent less likely to be in school during the same period. The both estimated

coefficients are significant at the 10 percent level.

Among the boys, we find a strong negative impact of the post-mortality variable on

schooling (-0.258, t-stat = -3.62), especially among the poor. The result indicates that boys in the

23

poor households are 25.8 percent less likely to be in school for some time after a death of a

working-age member. This could be because boys are replacing the loss of family labor in farm

production or other income generating activities, or that school fees can no longer be afforded.

Even among the less-poor households, we find a negative yet much smaller and weaker impact of

the post mortality variable on boys’ schooling (-0.062 with t-stat=-1.64).

Fixed effects models

Next, we estimate various fixed effects models based on Equation 3 to determine the

effects of adult death on child school attendance (Table 6). Again, we estimate the model

separately for the poor and non-poor (columns B and C). We present the results from conditional

logit models with household-fixed effects and with child-fixed effects. Note, however, that only

observations with variations in the dependent variable within a group can be included for

estimations in conditional logit models. This is why there are fewer observations in the

conditional logit models. Especially when we stratify the sample by initial wealth, the number of

observations declines significantly and many estimators are no longer precisely identified. Thus,

we face a trade-off between the consistency and efficiency. Because of this trade-off, we use the

results in Table 6 mainly as a check for robustness of the results in Table 4 and 5.

The results in Table 6 indicate that even after controlling for the child-fixed effects, the

lagged HIV prevalence rate retains its significantly negative influence on schooling among the

poor. Thus, the lagged HIV-prevalence rate is not only representing the fixed relationship

between the lagged HIV-prevalence rate and schooling but also a relationship between a change in

the lagged HIV-prevalence rate and a change in schooling. We do not find any significant

influence of the lagged HIV-prevalence rates on school attendance among the non-poor. 24

Turning to the effects of a working-age adult death in the household, we find that children

in the relatively poor households are less likely to be in school in the one to two year period before

their households experience a working-age death. The household fixed effects models also

indicate that within the household, there is no significant difference in schooling between

household heads’ own children and grandchildren, nephews, and nieces. The only group of

children who are less likely to be in school compared with the household heads’ own children in

the same household is “non-relative” children who have distant relationships or no relationship

with the household heads. These results are similar to those found by Case, Paxson, and

Aleidinger (2002) and suggest that children are treated equally as long as they are living with their

close relatives. Unfortunately, because we do not know the orphan-status of all sampled children,

we cannot determine whether orphans are treated differently than non-orphans in their relatives’

households. However, there is likely to be some overlap between the “non-relative” category and

those children in the sample who are orphans.

6. Conclusions

A rapidly increasing mortality rate among adults due to the AIDS epidemic in sub-Saharan

Africa has raised concerns about intergenerational effects, including child education. Using a

panel of 1,422 households in rural Kenya, we first estimated the determinants of primary school-

aged children experiencing working-age adult mortality. We find a high correlation between

lagged HIV-prevalence at nearby urban sentinel surveillance sites and the probability that a child

aged 7 to 14 years experiences the death of a working-age adult in his or her household. The

lagged HIV prevalence rate is negatively correlated with primary school attendance, especially

among girls, and in relatively poor households (those whose initial asset levels at the time of the 25

first survey in 1997 were below the median) even after controlling for working-age adult mortality.

These findings may indicate that there are important community impacts of the HIV/AIDS

epidemic on schooling among the poor, although it cannot be ruled out that this variable is picking

up unobserved and unrelated factors correlated with these changes in HIV prevalence over time.

We also find that children, especially girls, in relatively poor households are less likely to

be in school in the period prior to the adult death in their household compared with children in

unafflicted households. In the 1-2 year period directly after the death of an adult, there is a

significant decline in the probability that boys in relatively poor households attend school. The

fact that we find larger impacts among the poor could be because households closer to the edge of

economic survival are forced to take more extreme measures to adjust to major shocks to their

livelihoods, even at the expense of long-run human capital development, than their neighbors.

Because they have fewer options, poor households appear more likely to reallocate their children’s

time from schooling to care-giving for sick adults or to providing labor to compensate for the lost

labor of the sick adult. However, because of this study spans only a few years, it is unclear

whether the effects on schooling as measured in this study are long-term or only temporary.

Because of limited capacity of local hospitals compared with the overwhelming number of

AIDS patients, most AIDS patients are taken care of at their homes in rural areas. Although

home-based care should be promoted to ease the burden on the medical system in rural areas,

children appear to be bearing part of the burden of taking care of the sick. Policies to reduce the

burden of taking care of the sick at home, such as improved community health care systems, may

have an added advantage of helping afflicted households keep their children in school.

In Kenya, primary school enrollment rates are high compared with neighboring countries.

Thus, it is quite possible that we would find much larger impacts of the working-age adult 26

mortality on the primary school enrollment in other countries. There is a need to deepen our

understanding on the impacts of adult mortality on child schooling in countries that are suffering

from the AIDS epidemic so that governments and donor institutions can determine the best use of

limited financial and human resources to mitigate the impacts of HIV/AIDS. Otherwise, there is

the risk that the AIDS epidemic may produce as yet unanticipated intergenerational consequences

on human capital development that might have been mitigated if understood and addressed earlier.

27

References

Ainsworth, M. and J. Dayton. 2003. “The Impact of the AIDS Epidemic on the Health of Older Persons in Northwestern Tanzania.” World Development 31 (1): 131-148. Ainsworth, M., K. Beegle, and G. Koda. 2002. “The impact of adult mortality on primary school enrollment in northwestern Tanzania.” African Region Human Development Working Paper Series, no. 23961, The World Bank, Washington DC. Ainsworth, M., and D. Filmer. 2002. “Poverty, AIDS and children’s schooling: a targeting dilemma.” World Bank Policy Research Working Paper 2885, World Bank, Washington D.C. Ainsworth, M. and I. Semali. 1998. “Who is most likely dies of AIDS? Socioeconomic correlates of adult deaths in Kagera Region, Tanzania.” In Confronting AIDS: Evidence from the developing world, ed. M. Ainsworth, L. Fransen, and M. Over. Brussels: European Commission. Ainsworth, M., Ghosh, S., Semali, I. 1995. “The Impact of adult deaths on household composition in Kagera Region, Tanzania.” Unpublished manuscript, Policy Research Department, World Bank, Washington, D.C. Alderman, H., J. Behrman, H. Kohler, J. Maluccio, and S. Watkins. 2001. “Attrition in longitudinal household survey data: some tests for three developing country samples.” Demographic Research, vol. 5: 78-124. Barnett, T., and P. Blaikie. 1992. AIDS in Africa: its present and future impact, Belhaven Press. Barro, Robert. 1991. “Economic growth in a cross section of countries.” Quarterly Journal of Economics, vol.106: 107-158. Beegle, K. 2003. “Labor Effects of Adult Mortality in Tanzanian Households.” World Bank Policy Research Working Paper 3062, World Bank, Washington D.C. Bell, C., S. Devarajan, and H. Gersbach. 2003. “Long Run Economic Costs of AIDS: Theory and Application to South Africa.” Report posted on Development Gateway; url: http://www.developmentgateway.org/node/130640/dg-contribute/item-detail?item%5fid=341185&version%5fid=204458&from=alert Carswell, J.W. G. Lloyd, and J. Howells. 1989. “Prevalence of HIV-1 in East African lorry drivers.” AIDS, 3:759-761. Case, A., C. Paxson, and J. Aleidinger. 2002. “Orphans in Africa.” NBER Working Paper Series No 9213, National Bureau of Economic Research, Cambridge, MA. Fitzgerald, J., P. Gottschalk, and R. Moffitt. 1998. “An analysis of sample attrition in panel data.” The Journal of Human Resources, vol. 33: 251-299. 28

Fox, M., S. Rosen, W. MacLeod, M. Wasunna, M. Bii, G. Foglia, and J. Simon. 2003. “The Impact of HIV/AIDS on Labor Productivity in Kenya.” Draft discussion paper, Center for International Health and Development, Boston University, Boston. Gabre-Madhin, E., and B.F. Johnston. 2002. “Accelerating Africa’s Structural Transformation: Lessons from East Asia.” In Perspectives on Agricultural Transformation: A View from Africa, ed. T.S. Jayne, G. Argwings-Kodhek, and I. Minde, New York: Nova Science. Gachuhi, D. 1999. “The Impact of HIV/AIDS On Education Systems in The Eastern and Southern Africa Region and the Response of Education Systems to HIV/AIDS: Life Skills Programmes.” Report for UNICEF, Eastern and Southern Africa Regional Office, Nairobi. Glewwe, P. 2002. “Schools and skills in developing countries: education policies and socioeconomic outcomes.” Journal of Economic Literature, vol.40: 436-482. Gregson, S., H. Waddell, and S. Chandiwana. 2001. “School education and HIV control in Sub-Saharan Africa: from discord to harmony?” Journal of International Development, vol.13: 467-85. Guest, E. 2001. Children of AIDS: Africa’s Orphan Crisis. Pluto Press, London. Johnston, B.F. and P. Kilby.1975. Agriculture and Structural Transformation: Economic Strategies in Late-Developing Countries. New York: Oxford University Press. Lundberg, M., M. Over, and P. Mujinja. 2000. “Sources of financial assistance for households suffering an adult death in Kagera.” South African Journal of Economics, 68: 947-984. Mankiw, N., Romer, D. and Weil, D.N. 1992. “A contribution to the empirics of economic growth.” Quarterly Journal of Economics, vol.107: 407-437. Mbugua, G.G., L.N. Muthami, C.W. Mutura, et al. 1995. “Epidemiology of HIV infection among long distance truck drivers in Kenya.” East African Medical Journal, vol.72: 515-518. McPherson, M.F. (2001) “HIV/AIDS Impacts on Capacity Deepening and Economic Growth” Paper presented at The Brookings Institution, Washington D.C., 28th June NASCOP (National AIDS and STDs Control Programme). 2001. Estimating National HIV Prevalence in Kenya from Sentinel Surveillance Data, NASCOP, Nairobi, Kenya, June report. Nyambedha, E.O., S. Wandibba, and J. Aagaard-Hansen. 2001. “Policy implications of the inadequate support systems for orphans in Western Kenya.” Health Policy, vol.58: 83-96. Nzyuko, S., P. Lurie, W. McFarland, W. Leyden, D. Nyamwaya, and J.S. Mandel. 1997. “Adolescent sexual behavior along the Trans-Africa Highway in Kenya.” AIDS, vol.11: S21-S26. 29

Opiyo, P. 2001. HIV/AIDS, Gender and Livelihood in Siaya District, Kenya: An Analysis of AIDS Impact on Rural Households. MSc. Thesis, Wageningen University. Rugalema, G. 1999. HIV/AIDS and the Commercial Agricultural Sector of Kenya. Report. FAO/UNDP, Rome. Strauss, J., and D. Thomas. 1995. “Human resources: empirical modeling of household and family decisions.” In Handbook of Development Economics, vol.3A, ed. J. Behman and T.N. Srivasan North-Holland, Amsterdam. Todd, J. 2003. “Age Patterns and Trends in HIV-Positive and HIV-Negative Mortality Rate Ratios.” Paper presented at the Scientific Meeting on Empirical Evidence for the Demographic and Socio-Economic Impact of AIDS 26-28 March 2003, University of Natal, Durban, South Africa. UNICEF. 1999. Children Orphaned by AIDS, UNICEF, Geneva. United Nations. 2000. Millennium summit goals, New York: United Nations. USAID. 2002. Children on the Brink 2002. USAID, Washington D.C. Yamano, T., and T.S. Jayne. 2004. “Measuring the impacts of working-age mortality on rural households in Kenya.” forthcoming, World Development. Withworth, J., L. Shafer C. Mahe and L. Van der Paal. 2003. “Survival since infection and its relation to background mortality n Masaka general population.” Paper presented at the Scientific Meeting on Empirical Evidence for the Demographic and Socio-Economic Impact of AIDS, 26-28 March 2003, University of Natal, Durban, South Africa. World Bank. 1995. World Development Report 1994/1995, New York: Oxford University Press. World Bank. 1999. Confronting AIDS: Public priorities in a global epidemic. Revised edition. New York: Oxford University Press. World Bank. 2002. World Development Report 2001/2002. New York: Oxford University Press.

30

Prob(S=1): Probability of Being in School

Non-Poor

Time Death

1

0

ba

After death period Before death period

Illness

Poor possible

recovery

Figure 1. Impact of Working-age Adult Mortality: Hypotheses Notes: (a) hypothesized decline in probability of school attendance due to greater opportunity cost of children’s time and financial resources. (b) possible recovery period due to reduced demands on children’s time for care giving.

31

Table 1. Working-Age a Adult Mortality and School Enrollment in Rural Kenya

Sampled Households % of Children aged 7-14 in School

Province District

in 1997 & 2000b

in 1997, 2000, & 2002c

Households incurring

working-age mortality, 1997-2002

2000

2002

HIV prevalence at urban sentinel

sites in 1990-94d

(A) (B) (C) (D) (E) (F) - % - - % - - % - - % - Coastal 88 71 11.4 59.0 67.5 13.2 Eastern 233 215 6.4 87.1 94.0 6.3 Nyanza 262 245 17.6 89.0 88.2 12.6 Kisumu/Siaya 177 176 23.7 88.4 84.8 21.6 Kisii 85 69 4.7 90.2 95.4 3.6 Western 290 272 7.2 80.9 82.4 15.8 Central 174 164 5.2 92.3 88.5 9.6 Rift Valley 375 299 5.3 87.8 88.0 12.3 Total 1,422 1,266 8.5 89.7 88.3 Source: Tegemeo Institute/Egerton University Agricultural Monitoring and Policy Analysis Household Surveys in 1997 and 2000. National AIDS and STDs Control Programme (NASCOP, 2001) Note: (a) Working-age is defined as 15-49 for women and 15-54 form men. (b) Sample is restricted to households that were interviewed in the 1997 and 2000 surveys. (c) Sample is restricted to households that were interviewed in the 1997, 2000, and 2002 surveys. (d) The average percentage of pregnant women who visited the urban-sentinel-surveillance sites and tested HIV positive in 1990-1994. Data are taken from 11 urban sentinel surveillance sites (NASCOP, 2001).

32

Table 2. Primary School Enrollment by Working-age Adult Mortality

Did household incur Working-age Mortality? NO YES between 1997-2000 YES between 2000-2002

2000 2002 2000 2002 2000 2002

(A) (B) (C) (D) (E) (F) All Children Enrollment (%) 90.1 88.8 91.1 81.6 70.0 89.6 Grade (years) 3.50 2.88 3.58 2.86 2.78 2.38 Age (years) 10.7 10.8 10.7 10.9 10.5 10.7 Number 2,358 1,934 157 125 50 48

Poor Enrollment (%) 88.4 86.9 86.4 72.1 61.3 83.3 Grade (years) 3.25 2.66 3.62 2.44 2.03 1.90 Age (years) 10.6 10.7 10.8 10.7 10.1 10.5 Number 1,170 951 66 43 31 30

Non-Poor Enrollment (%) 91.8 90.6 94.5 86.6 84.2 100 Grade (years) 3.75 3.09 3.55 3.09 4.00 3.17 Age (years) 10.8 10.8 10.6 10.9 11.1 10.9 Number 1,188 983 91 82 19 18

Source: Tegemeo Institute (Nairobi)/Michigan State University Agricultural Monitoring and Policy Analysis Household Surveys in 1997, 2000, and 2002.

33

Table 3. Determinants of Children Experiencing Working-age Adult Mortality in Their Household, 2000 Survey Sample (Probita)

Working-age adult mortality in 1997-2002: Prob (D1997-2000=1) (A) (B)

HIV Prevalence rate Ratio of HIV+ pregnant women 0.487 In 1990-94 (5.22)** Attrition in 2002 Attrition Dummy (=1) 0.027 0.021 (2.15)* (1.75) Province Dummies Nyanza province 0.085 0.050 (5.14)** (3.32)** Coastal province 0.020 0.026 (1.00) (1.29) Western province -0.013 -0.002 (0.94) (0.15) Central province -0.035 -0.007 (2.21)* (0.34) Rift Valley province -0.024 -0.017 (2.11)* (1.51) Household Characteristics Max. years of male schooling -0.001 0.000 (0.46) (0.02) Max. years of female schooling -0.001 -0.001 (0.98) (1.16) Female headed (=1) 0.001 -0.005 (0.05) (0.29) Polygamous household (=1) -0.023 -0.025 (2.09)* (2.48)* Number of male adults 0.014 0.012 (3.58)** (3.32)** Number of female adults -0.003 -0.002 (0.57) (0.44) Land tenure (=1) 0.012 0.013 (1.41) (1.48) ln (Landholding size in acres) 0.011 0.009 (2.01)* (1.83) ln (Asset value, Shillings) 0.001 0.001 (0.60) (0.77) Distance to Bus Stop (km) -0.004 -0.005 (1.87) (2.34)* Distance to Piped Water (km) -0.000 0.000 (0.02) (0.29) Child Characteristics Age in years 0.007 0.008 (0.34) (0.41) Age squared -0.000 -0.000 (0.45) (0.50) Girl (=1) -0.001 -0.001 (0.09) (0.13) Grand child, Nephew, Niece (=1) 0.019 0.009 (1.78) (0.95) Non Relative (=1) 0.018 0.012 (0.77) (0.54) Number of children 2,565

Note: Numbers in parentheses are absolute z-scores. ** indicates 1 percent significance level; * indicates 5 percent significance level. (a) Estimated coefficients are marginal changes in probability.

34

Table 4. The Impact of Prime-age Adult Mortality on School Attendance, Pooled 2000 and 2002 Data (Probita), Dependent Variable: Prob(Sit =1)

All Poor Non-Poor (A) (B) (C)

HIV Prevalence at surveillance sites Ratio of HIV+ pregnant women -0.215 -0.310 -0.141 of 1990-94 at 2000, 1992-96 at 2002 (2.94)** (3.19)** (1.23) Working-age Adult Mortality

Dt: Working-age Adult Death at t 0.018 0.015 0.024 (1.35) (0.82) (1.68)

Dt+1: Pre-death Period -0.099 -0.143 -0.006 (2.68)** (2.48)* (0.19)

Dt-1: Post-death Period -0.060 -0.106 -0.024 (2.70)** (2.50)* (1.16) Attrition in 2002 Attrition Dummy (=1) 0.011 -0.015 0.020 (0.95) (0.78) (1.47) Child Characteristics

Age in years 0.026 0.015 0.027 (1.85) (0.73) (1.70)

Age squared -0.002 -0.002 -0.002 (2.68)** (1.62) (2.16)*

Girl (=1) 0.001 -0.004 0.008 (0.19) (0.50) (1.07)

Grade completed, splined at 3rd grade 0.073 0.105 0.044 (21.27)** (16.76)** (12.47)**

Grade completed, 3rd and higher -0.012 -0.022 -0.005 (3.89)** (4.87)** (1.30)

Grand child, Nephew, Niece (=1) -0.002 0.012 -0.023 (0.21) (1.10) (2.22)*

Non Relative (=1) -0.341 -0.294 -0.340 (12.16)** (5.92)** (10.67)** Household Characteristics

Max. years of male schooling -0.001 0.000 -0.003 (1.48) (0.05) (2.43)*

Max. years of female schooling 0.004 0.005 0.002 (3.68)** (3.41)** (1.44)

Female headed (=1) 0.022 0.036 -0.022 (1.53) (2.23)* (0.73)

Polygamous household (=1) 0.011 0.013 0.009 (1.33) (1.13) (0.87)

Number of male adults -0.001 -0.001 0.000 (0.17) (0.20) (0.02)

Number of female adults -0.005 -0.010 -0.001 (1.71) (1.91) (0.44)

Land tenure (=1) 0.004 -0.005 0.002 (0.63) (0.46) (0.19)

ln (Landholding size in acres) -0.004 0.009 -0.007 (0.92) (1.26) (1.59)

ln (Asset value, Shillings) 0.003 0.004 0.003 (2.51)* (2.22)* (1.20)

Distance to Bus Stop (km) -0.001 -0.000 -0.001 (1.87) (0.22) (2.83)**

Distance to Piped Water (km) 0.002 0.003 0.002 (1.66) (1.45) (1.37)

Year 2002 0.018 0.026 0.017 (2.77)** (2.72)** (2.15)* Number of children 4,672 2,291 2,381

Note: Numbers in parentheses are absolute z-scores. ** indicates 1 percent significance level; * indicates 5 percent significance level. (a) Estimated coefficients are marginal changes in probability.

35

Table 5. The Impact of Prime-age Adult Mortality on Child School Attendance by Gender, Pooled 2000 and 2002 Data (Probita), Dependent Variable: Prob(Sit =1)

All Poor Non-Poor Boys Girls Boys Girls Boys Girls

(A) (B) (C) (D) (E) (F) HIV Prevalence at surveillance sites Ratio of HIV+ pregnant women -0.230 -0.131 -0.186 -0.369 -0.187 0.036 of 1990-94 at 2000, 1992-96 at 2002 (2.31)* (1.27) (1.65) (2.64)** (1.22) (0.23) Working-age Adult Mortality

Dt: Working-age Adult Death at t -0.003 0.032 -0.025 0.033 0.018 0.026 (0.15) (1.86) (0.95) (1.49) (0.97) (1.41)

Dt+1: Pre-death Period -0.082 -0.123 -0.124 -0.151 0.026 -0.105 (1.65) (2.22)* (1.80) (1.77) (0.80) (1.41)

Dt-1: Post-death Period -0.145 0.009 -0.258 0.015 -0.062 0.009 (3.72)** (0.39) (3.62)** (0.34) (1.64) (0.44) Attrition in 2002 Attrition Dummy (=1) 0.015 0.007 0.003 -0.027 0.016 0.014 (0.99) (0.43) (0.16) (0.93) (0.90) (0.79) Child Characteristics

Age in years 0.035 0.006 0.028 -0.009 0.034 0.020 (1.94) (0.31) (1.28) (0.30) (1.56) (0.94)

Age squared -0.002 -0.001 -0.002 -0.001 -0.002 -0.001 (2.40)* (0.94) (1.85) (0.39) (1.79) (1.23)

Grade completed, splined at 3rd grade 0.072 0.070 0.088 0.107 0.044 0.036 (15.93)** (13.85)** (11.99)** (11.27)** (9.41)** (7.67)**

Grade completed, 3rd and higher -0.017 -0.011 -0.023 -0.020 -0.009 -0.004 (4.17)** (2.38)* (4.52)** (3.04)** (1.93) (0.78)

Grand child, Nephew, Niece (=1) -0.011 0.005 -0.011 0.029 -0.020 -0.025 (1.11) (0.43) (0.88) (1.91) (1.45) (1.81)

Non Relative (=1) -0.549 -0.148 -0.617 -0.046 -0.504 -0.164 (11.84)** (4.46)** (6.66)** (0.94) (9.31)** (4.58)** Household Characteristics

Max. years of male schooling -0.002 0.000 0.001 0.000 -0.004 -0.001 (1.97)* (0.02) (0.47) (0.01) (2.77)** (0.49)

Max. years of female schooling 0.003 0.004 0.002 0.007 0.003 0.000 (2.41)* (2.68)** (1.41) (3.22)** (1.84) (0.13)

Female headed (=1) 0.008 0.034 0.020 0.049 -0.056 -0.004 (0.41) (1.68) (1.11) (2.11)* (1.15) (0.12)

Polygamous household (=1) 0.007 0.014 -0.012 0.033 0.023 -0.001 (0.60) (1.17) (0.78) (2.05)* (1.69) (0.09)

Number of male adults 0.004 -0.006 0.004 -0.005 0.004 -0.003 (0.93) (1.33) (0.75) (0.72) (0.86) (0.69)

Number of female adults -0.004 -0.007 -0.008 -0.012 -0.000 -0.004 (1.09) (1.55) (1.56) (1.45) (0.02) (0.85)

Land tenure (=1) 0.007 -0.001 -0.000 -0.011 0.002 -0.002 (0.75) (0.09) (0.01) (0.69) (0.15) (0.19)

ln (Landholding size, acres) -0.003 -0.001 0.003 0.011 -0.005 -0.006 (0.47) (0.18) (0.43) (1.18) (0.92) (1.04)

ln (Asset value, Shillings) 0.003 0.003 0.003 0.002 0.000 0.007 (1.73) (1.79) (1.89) (1.00) (0.09) (1.83)

Distance to Bus Stop (km) -0.000 -0.001 0.000 -0.000 -0.001 -0.001 (1.25) (1.56) (0.83) (0.63) (2.78)** (2.02)*

Distance to Piped Water (km) 0.004 0.000 0.002 0.004 0.006 -0.001 (2.43)* (0.03) (0.82) (1.47) (2.83)** (0.67)

Year 2002 0.022 0.011 0.026 0.023 0.021 0.004 (2.45)* (1.11) (2.35)* (1.63) (2.00)* (0.42) Number of children 2,421 2,251 1,197 1,094 1,224 1,157

Note: Numbers in parentheses are absolute z-scores. ** indicates 1 percent significance level; * indicates 5 percent significance level. (a) Estimated coefficients are marginal changes in probability.

36

Table 6. The Impact of Prime-age Adult Mortality on Child School Attendance, Controlling for Fixed Effects, Pooled 2000 and 2002 Data (Conditional Logit)

All Poor Non-Poor HH FE Child FE HH FE Child FE HH FE Child FE

(A) (B) (C) (D) (E) (F) HIV Prevalence rate Ratio of HIV+ pregnant women -12.82 -47.89 -28.63 -50.02 21.90 -31.65 of 1990-94 at 2000, 1992-96 at 2002 (1.02) (2.59)** (1.81) (2.16)* (1.02) (0.83) Working-age Adult Mortality

Dt+1: Pre-death Period -1.745 -1.770 -1.783 -1.252 -32.51 -38.88 (2.09)* (1.58) (1.94) (1.05) (0.00) (0.00)

Dt-1: Post-death Period -1.075 -0.958 -0.702 -0.061 -1.369 -38.84 (1.66) (1.22) (0.88) (0.06) (1.28) (0.00) Child Characteristics

Age in years 0.766 1.241 0.510 2.195 1.174 -0.196 (2.04)* (1.78) (1.00) (2.21)* (2.04)* (0.16)

Age squared -0.043 -0.058 -0.031 -0.108 -0.060 0.003 (2.38)* (1.80) (1.26) (2.36)* (2.20)* (0.05)

Girl (=1) -0.015 0.137 -0.199 (0.09) (0.59) (0.71)

Grade completed, splined at 3rd 1.342 0.599 1.523 0.887 1.136 0.664 (11.7)** (3.60)** (9.65)** (3.48)** (6.66)** (2.30)*

Grade completed, 3rd and higher -0.394 -0.796 -0.512 -0.607 -0.305 -1.543 (4.21)** (2.87)** (4.08)** (1.83) (2.05)* (2.17)*

Grand child, Nephew, Niece (=1) 0.028 0.369 -0.477 (0.07) (0.74) (0.77)

Non Relative (=1) -2.883 -2.088 -3.366 (6.71)** (3.21)** (5.75)**

Year 2002 (=1) 0.384 0.916 0.509 0.912 -0.034 1.108 (1.47) (1.74) (1.46) (1.35) (0.08) (1.10) Number of observations 1,396 270 792 178 604 92 Note: Numbers in parentheses are absolute z-scores. ** indicates 5 percent significance level; * indicates 10 percent significance level.

37

Appendix Table A1. Descriptive Statistics by Gender, Pooled 2000 and 2002 Survey Data Mean Std. Error Min Max Schooling School Attendance (=1) 0.891 0.311 0 1 HIV Prevalence rate Ratio of HIV+ pregnant women 0.129 0.053 0.036 0.250 Working-age Adult Mortality

Dt: Working-age Adult Death at t (=1) 0.043 0.202 0 1 Dt+1: Pre-death Period (=1) 0.009 0.097 0 1 Dt-1: Post-death Period (=1) 0.027 0.161 0 1

Attrition in 2002 Attrition Dummy (=1) 0.074 0.262 0 1 Child Characteristics

Age in years 10.71 2.247 7 14 Girl (=1) 0.482 0.500 0 1 Grade completed 3.209 2.168 0 7 Grand child, nephew, niece (=1) 0.226 0.418 0 1 No relative child (=1) 0.040 0.196 0 1

Household Characteristics Max. years of male schooling 8.040 4.473 0 18 Max. years of female schooling 7.355 3.992 0 17 Female headed (=1) 0.057 0.232 0 1 Polygamous household (=1) 0.136 0.342 0 1 Number of male adults 1.824 1.214 0 8 Number of female adults 1.850 1.171 0 8 Land tenure (=1) 0.383 0.486 0 1 ln (Landholding size in acres) 1.511 0.902 0 4.615 ln (Asset value, Shillings) 5.695 4.359 0 15.85 Distance to Bus Stop (km) 2.313 2.626 0 20 Distance to Piped Water (km) 9.054 11.94 0 70 Year 2002 (=1) 0.451 0.498 0 1 Number of observations 4,672 Source: Tegemeo Institute (Nairobi)/Michigan State University Agricultural Monitoring and Policy Analysis Household Surveys in 1997, 2000, and 2002.

38

39

Appendix Table A2. Three-Year Average HIV-Prevalence Rates Among Pregnant Women At Urban Sentinel Surveillance Sites City 1990-94 1992-96 1995-99 2000 only Busia 20.4 25.0 28.6 22.0 Kakamega 11.2 12.0 12.0 12.0 Kisii 3.6 6.4 12.6 16.0 Kisumu 21.6 24.4 28.2 35.0 Kitale 9.8 12.4 12.6 17.0 Kitui 7.2 7.6 6.6 14.0 Meru 5.3 9.5 18.4 35.0 Mombasa 13.2 13.4 15.3 12.0 Nairobi 14.3 17.6 20.0 n.a. Nakuru 14.8 15.7 22.0 11.0 Nyeri 5.0 6.8 11.0 14.0 Source: NASCOP (2001).